In: Stilla U, Rottensteiner F, Hinz S (Eds) CMRT05. IAPRS, Vol. XXXVI, Part 3/W24 --- Vienna, Austria, August 29-30, 2005
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
MOTION DETECTION BY CLASSIFICATION OF LOCAL STRUCTURES IN
AIRBORNE THERMAL VIDEOS
M. Kirchhof, E. Michaelsen, K. Jäger
FGAN-FOM Research Institute for Optronics and Pattern Recognition,
76275 Ettlingen, Germany - kirchhof@fom.fgan.de
KEY WORDS: Detection of moving objects, classification, sensor pose estimation
ABSTRACT:
In this paper we present a method to efficiently detect moving objects namely vehicles in airborne thermal videos. The motion of the
sensor is estimated from the optical flow using projective planar homographies as transformation model. A three level classification
process is proposed: On the first level we extract interest location applying the Foerstner - operator. These are subject to the second
finer level of classification. Here we distinguish four classes: 1. Vehicles cues; 2. L-junctions and other proper fixed structure 3. Tjunctions and other risky fixed structure. 4. A rejection class containing all other locations. This classification is based on local
features in the single images. Only structures from the L-junctions class are traced as correspondences through subsequent frames.
Based on these the global optical flow is estimated that is caused by the platform movement. The flow is restricted to planar
projective homographies which highly reduces the computational time. This opens the way for the third classification. The vehicle
class is refined using motion as feature. Inconsistency with the estimated flow is a strong evidence for movement in the scene. This
is done by computing a difference image between two sequent frames transformed by a homography to be taken from the same
position. The difference images are pre-processed using vehicle properties and velocity.
use the pose estimation namely the estimated homography to
improve the detection of moving objects in the image.
1. INTRODUCTION
Detection of moving objects even if the observer moves is a
very important task for every creature in the world. Think on a
predator and a prey. In civilisations this is very important in
traffic and security aspects. Even for sensor calibration or
sensor pose estimation this is important because the movement
between two sequent frames is often so small, that it can not be
detected by RANSAC (Fischler & Bolles 1981) or similar
outlier detection algorithms. Thermal images provide unique
opportunities to detect vehicles and reveal their activity in
urban regions at any season in the year and at day or night.
Independent of the colour or type all vehicles appear similar
with respect to their size and outer conditions (Ernst 2005).
Depending on the resolution and aspect active vehicles or parts
of them may appear as hot spot (e.g. the exhaust). The exterior
of the body of fast vehicles takes the temperature of the
surrounding air. Often they will appear as cold spots on the
warmer road surface. On a sunny day high temperature
differences occur due to shadow and sun. So the appearance of
vehicles in such data varies strongly with the exterior condition.
Because the videos are taken from a moving platform the
optical flow caused by this movement has to be estimated in
order to distinguish the two types of motion in the images: Flow
caused by the platform motion versus motion caused by moving
objects in the scene.
Estimations of the flow caused by the sensor platform usually
assume the scene to be stationary. Therefore, vehicles may
cause substantial systematic error. If some of them move in the
same direction and cause correspondences with residual
movement below the threshold used for outlier decision they
may have a considerable impact and spoil the precision of the
estimation. For data from urban terrain with a lot of traffic this
is not unlikely. The main purpose of this contribution is to solve
both problems at the same time: Refine pose estimation by
reducing the systematical error caused by moving objects and
2. APRIORI CLASSIFICATION OF INTEREST
LOCATIONS
In order to exclude as many unreliable structures from the flow
estimation as possible we propose a two level classification of
image locations prior to it. The flow estimation is then followed
by a third step. We use only this third step of classification to
exclude moving objects from the flow estimation. In the first
level homogenous and boundary locations are detected and
excluded from further consideration. Only a few interest
locations remain for which the second level is performed which
consumes much more computation per location. Fig. 1 gives an
overview of the structure of our classification hierarchy. It is
described in more detail in the sections below.
Figure 1. Classification hierarchy for image locations
2.1 First Level: Interest Locations, Boundary Locations
and Homogenous Locations
It is not possible to localize correspondence between different
frames if the image is homogenous in that location. If an edge
or line structure is present at a location in the 2-d image array
there may still be an aperture problem. Secure correspondence
can only be obtained at locations where a corner, crossing or
spot is present. It is proposed to use the averaged tensor product
of the gradient g of the grey-values (Förstner, 1994)
63
CMRT05: Object Extraction for 3D City Models, Road Databases, and Traffic Monitoring - Concepts, Algorithms, and Evaluation
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
⎛ g
∇ g ∇ g T = ⎜⎜
⎝ gygx
2
x
gxgy ⎞
⎟
g y2 ⎟⎠
radii 1≤d≤dmax and obtain several functions over the interval
[0,2π). Figs. 3 display such sampling in the upper left position
under ‘Original’ whereby the functions are appended one after
the other so that the total domain is [0,dmax2π). This function is
normalized (byte to one) and average-free so that it takes
positive and negative values.
As proper rotation invariant features we chose the Fourier
power coefficients of each function. The first ten coefficients
i.e. dmax10 features are used in a nearest neighbour classifier. To
obtain a consistent metric we normalised the features to be in
the interval [0,1]. The feature vectors are displayed in the Figs.
3 upper right under ‘Discrete Fourier transform’. Also the greyvalues around the interest locations are shown in a window of
(2d max)2 pixel size (dmax was chosen to be 7 here). The lower
right part indicates the corresponding location in the image
using a black arrow.
(1)
where x and y indicate the directions in the image. The discrete
version of this matrix is obtained by convolution of the original
image with three masks successively (two for the directional
derivatives with inherent smoothing and one Gaussian for
averaging the squared gradient). For better precision we
recommend to use an hourglass like filter in the last averaging
step, oriented on the gradient direction (Köthe, 2003). We do
not further treat the interest-operator in this contribution.
Note that the remaining pixels have both eigenvalues
significantly non-zero and are called interest pixels from this
on. For image materials like the one presented in Section 3
around one ‰ of the pixels are classified as interest pixels. For
further reduction of computational complexity without losing
precision of the result we perform non maximum suppression
using the eight pixel neighbourhood.
2.2 Second Level: Spots, Corners, T-Junctions and other
Structure
As already proposed by Förstner (1994) the pixels around the
interest pixels class are grouped in clusters according to
proximity and local paraboloids are fitted to these clusters to
determine a unique location for each such cluster with sub-pixel
accuracy. The result is the set of interest locations.
In order to make sure that the interest locations are distributed
over the whole image and do not cluster too much in densely
structured areas, we set an equally spaced grid (e.g. ten by ten)
over the image. Inside of each sub-image only a limited number
of interest locations (e.g. maximal fifteen) is accepted. If there
are more interest locations available only those with the best
value for the structure tensor will be accepted.
Felsberg & Sommer (2001) theoretically derive the
recommendation of using polar coordinates once interest
locations have been detected by local energy maximizations
like the structure tensor operator presented above. As practical
consequence there is a research line particularly concerning
structure based correspondence evaluation for stereo based on
this decomposition of local structure around interest locations
(Krüger & Felsberg 2004).
Figure 3. Example for an interest location of the L-class For each radius
more than one of the lower frequencies appears strongly. For the larger
radii this gets stronger
The following three classes were distinguished:
Class 1 – L-class: The main purpose of the classification is to
distinguish among the interest locations those that may
contribute reliably to the homography estimation from those
that do not. We call this class the L-class because it contains
things like the corners of terrain regions of different
temperature. In urban terrain this includes also building
vertices. So also Y-junctions belong to this class.
d
Class 2 – T-class: One source of systematic error for the
homography estimation are image structures that result from
partial occlusion. We call this class the T-class because
typically the occluding part crosses an occluded boundary like a
T. Such structure will often not be detected as outlier by the
sub-sequent RANSAC analysis described in Sect. 3. because
the systematic error is below the thresshold. We include also
other unreliable structure like those locations that do not
provide robust correspondence – e.g. for lack of curvature.
Figure 2. The circular grey-level function: Left pixel grid with interest
location and circle around it; right grey–level function along this path
Following this we define a cyclic one-dimensional function gd
along a circular path of radius d around each interest location
containing the grey-values as function on the interval [0,2π).
Fig. 2 shows the principle. Note that while the grey-values are
associated to discrete pixel positions the centre of the circle is
located at the interest location which is determined with subpixel accuracy. In order to avoid strong dependence of the
results on the parameter d we repeat the circular sampling at all
Class 3 – vehicle cues: Vehicles often move and thus violate
the assumptions made for the homography estimations. If their
movement is small they may not be detected as outliers by the
RANSAC search just like the T-class objects and cause
systematic deviation of the estimation. On this stage we can
only detect vehicles that appear sufficiently small, so that they
64
In: Stilla U, Rottensteiner F, Hinz S (Eds) CMRT05. IAPRS, Vol. XXXVI, Part 3/W24 --- Vienna, Austria, August 29-30, 2005
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
build pairs of solutions with consistent normal vector of the
scene plain and compute the back-projection-error as last
criterion. The constraint that all correspondences are located on
the same 3D-plane reduces the problem from 24 to 14
parameters describing the homographies uniquely. To further
improve the homography estimation we only use those view
correspondence-triple that have –after initialising over the first
triple – the first interest point classified as L-class after the last
step of classification.
Figs. 5 and 6 show exemplarily that prior classification of the
interest locations and restriction of the input to the estimation
process only to the most reliable class – the L-class - leads to
high improvement of the flow estimation. The first image was
transformed with the computed homography to match the same
sensor position as the second but to a sequent time period. Then
the absolute differences between the transformed first image
and the second image of a pair is computed and coded in grey
in Fig. 6.
can be discriminated as being spot-shaped by our features
within the radius dmax.. Larger vehicles in the foreground will
often end up in the L-class, because they are to big to be
recognized by local features. But these are usually fast enough
to be recognized later as not consistent with the homography.
Fig 4. shows the result after the first two classification steps.
You can see many falsely detected vehicles because all spot
shaped objects were put in this class.
Figure 4. A priory classification of interest location: +=reject, T=Tclass, L=L-class, □=vehicle cue class
3. MOTION DETECTION
The main feature for the recognition of vehicles is their
movement in the scene. However movements in the aerial video
result from two sources, the vehicle movement and the
movement of the platform. We are interested in oblique views
from sufficiently high moving planes. The appropriate model
for the optical flow caused by platform movements is therefore
a planar projective homography (Hartley & Zisserman 2000)
which is estimated from the video itself.
Any correspondence trace of interest points not consistent with
this mapping is resulting from a moving vehicle with high
significance or objects far outside the main scene plain. The
homography estimation must be robust and precise. If the
precision is too bad many interest points may violate the
homography mapping and will falsely be detected as vehicle.
Figure 5. Homography differences without prior classification
3.1 Robust Homography estimation and difference image
We therefore basically use the RANSAC method (Fischler &
Bolles 1981). But there are some modifications and
improvements. At first we perform a standard adaptive
RANSAC with guided matching (Hartley & Zisserman 2000).
This means that the algorithm computes the number of samples
on its own. Based on the lowest outlier rate this is done by
assuring that with a probability of 99% one of the samples
consists of inliers only. Then a least square solution is
computed and the outliers are proved on this hypothesis. To
increase the precision by redundancy we compute three
homographies for each time step t, i.e. frame pairs (t,t+1),
(t+1,t+2) and (t,t+2). Taking only correspondences that were
tracked through all three images we perform bundle adjustment
for homographies (Kraus 1994, Hartley & Zisserman 2000)
using minimal parameterization. To this end we compute the
quadric decomposition of the homography as proposed by
Faugeras (1993). To select the right one of the two solutions we
Figure 6. Homography differences with prior classification
3.2 Segmentation and physical information
In the difference image we see different types of objects. There
are some systematical errors caused by the fact that the scene is
no texture on a plan but truly three-dimensional. The larger the
scene depth is the more systematical errors appear in the scene.
Mostly we see grey value differences caused by the motion of
objects like vehicles.
We have to decide between the two object classes: systematical
error from motion and from three-dimensional structure. For
65
CMRT05: Object Extraction for 3D City Models, Road Databases, and Traffic Monitoring - Concepts, Algorithms, and Evaluation
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
this decision we use some physical information about the
objects we like to detect. We can not compute the exact
velocity of a vehicle from one difference image but we can
compute the velocity relative to the width of the vehicle. Setting
a default velocity we can perform segmentation of the
difference image. The computed segments can be evaluated
using properties like roundness or eccentricity. Objects with a
distinct three dimensional structure – like buildings – cause
segments with high eccentricity which can easily be
distinguished from the others. Fig. 7 and 8. show the different
results after the segmentation and after using physical
information. The different segments are coded in multiple
colours.
3.4 Limitations
The first kind of limitations we have to take care of is shown in
Fig. 12. If we have to deal with many occlusions the detected
motion will often be smaller than the true and can not be
detected over the whole sequence. The second limitation we
have to take care of comes from the relative scene depth. Figure
9 shows the maximal relative scene depth dependent on the
velocity of the air vehicle under two different depression angles
and two different maximal errors in the image. If the scene
depth is too strong the system can be changed to work in the
motion detection step with half resolution. This is in general
only the case if the flight altitude is too low and the vehicles we
want to detect become very big. This means that the system also
works under worse conditions.
Figure 7 default segmentation result
Figure 9 maximal scene depths dependent on velocity
One different way to deal with the errors caused by scene depth
was presented in section 3.2. We have to remember that the
main scene depth comes from extended structure and therefore
lead to extended objects in the difference image. This is why in
Fig. 12 false alarms are only isolated single.
4. EXPERIMENTS AND CONCLUSION
4.1 The Data
Both example videos had been obtained with an AIM camera
with focal plane array sensitive in the mid thermal domain at 35µm. It was forward looking mounted to a helicopter platform.
They were taken during day-time, so that vehicles and
vegetation appear darker (i.e. colder) than the background. The
road surface is quite warm. One frame is displayed in Fig. 10.
This needs special care in order not to disturb the flow
estimation. Some vehicles on the right lanes exhibit hot
exhausts. Vehicles are of considerable different size.
One frame of the second sequence - called small road scene - is
shown in Fig. 11. In this scene we had to deal with very low
flight altitude which increases the scene depth and many
occlusions on the road. Even our reference classification - made
the human eye failed in the most frames of this sequence.
Fig. 12 shows the classification result of one frame from the
autobahn scene after using the movement feature to detect the
vehicles. Note that in contrast to the prior classification
presented in Fig. 4 almost no false alarm appears off the road.
On the other hand almost all vehicles are correctly marked now.
Figure 8 segmentation result with physical information
3.3 Motion detection in the context of interst location
The last step in our approach is the motion detection. In contrast
to other methods we analyse only the motion of our interst
locations. This highly reduces the computational time. We
analyse the region around our interst location in the difference
image and can set different thresholds for the size of the nearest
segment for our decision based on the a priori information.
The results are shown in Fig.12 and 13.
66
In: Stilla U, Rottensteiner F, Hinz S (Eds) CMRT05. IAPRS, Vol. XXXVI, Part 3/W24 --- Vienna, Austria, August 29-30, 2005
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
Figure 10. One frame of the autobahn scene (grey values have been
adapted for good visibility here)
Figure 12. Posterior classification for autobahn scene
Figure 11. One frame of the small road scene (grey values have been
adapted for good visibility here)
Figure 13. Posterior classification for small road scene
4.3 Discussion and Future Research Lines
4.2 Results
Learning samples –about ten for the first three classes- have
been taken from the first image of the 400 frames of the first
video. For the experiment we used then only the last 200 frames
of this video to ensure that we made no self-classification. For
the second experiment we used the same trainings data. The
prior classification may suffer from inadequate samples if the
parameters of the scene (daytime, season, wheather) or of the
camera change. On the other hand the posterior classification
opens the way for automatic adaptation of the learning example
set. We may use constantly appearing moving vehicles as new
learning examples for the vehicle class and remove those old
ones that could not be affirmed. Also we may use the residual
error after estimating the flow as criterion to assess old and new
learning samples for the L- and T-class. For experiments
following this line of research we need a larger data set.
Table 1 and 2 show the classification results of our approach on
two randomly picked frames of each scene as absolute values.
Moving objects
Non moving objects
Detected motion
No motion detected
38 …|… 39
1
...|... 5
0
...|... 3
1045 ...|... 995
Table 1: Detected motion in the autobahn scene
Moving objects
Non moving objects
Detected motion
No motion detected
2 ...|... 16
0
...|... 5?
0 ...|... 1
995 ...|... 887
Table 2: Detected motion in the small road scene
Here the non moving objects are all inters points from scene
fixed structure. You can see the relative results averaged over
ten randomly picked frames of each sequence in Table 3. Note
that for the small road scene even the human classification
failed so that some results are shown with a “?” to remind on
the inaccuracy.
Moving objects
Deviations in the optical flow from the proper homography may
not only result from movement but also from violation of the
planarity assumption. Actually, the terrain around the Autobahn
shown in Fig. 10 is not flat at all. Yet, in this example the flight
altitude is so high compared to differences in the terrain (scene
depth) and the time interval for the correspondences between
the frames so short, that such effects hardly matter. The 3d
structure can be understood as texture on the main plane.
Non moving objects
Detected motion
No motion detected
91.30% | 88 ±4%
0.30% | 0.8 ±0.7%
8.70% | 12 ±4%
99.70% | 99.2 ±0.7%
Table 3: Left: autobahn scene , right: small road scene
67
CMRT05: Object Extraction for 3D City Models, Road Databases, and Traffic Monitoring - Concepts, Algorithms, and Evaluation
¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯¯
Fischler, M. A., Bolles, R. C., 1981. Random sample
consensus: A paradigm for model fitting with applications to
image analysis and automated cartography. Comm. Assoc.
Comp. Mach., Vol. 24, pp. 381-395.
We could also show that for more difficult terrain like in Fig. 11
the systematical error from scene depth could be suppressed by
the segmentation described in section 3.2. Particularly, for low
flying platforms over urban terrain with high buildings,
however, they will. Albeit violating the homography transform
such non-planar structure flow must fulfil another weaker
constraint – the epipolar condition which is best captured in the
fundamental matrix. This can also be estimated from
correspondences (Hartley & Zisserman 2000). In such
situations an additional classification of interest locations is
recommended – consistent with the epipolar constraint versus
violating it. The latter must really be moving, while nothing can
be asserted about the former. Those may be moving in the
direction of the epipolar line. Further studies in this direction
are intended. They require appropriate example videos.
Foerstner, W. 1994. A Framework for Low Level Feature
Extraction. In: Eklundh J.-O. (ed.) Computer Vision – ECCV
94. vol. II, pp. 383-394.
Haag, M., Nagel, H.-H., 1999. Combination of edge element
and optical flow estimates for 3D-model based vehicle tracking
in traffic image sequences. IJCV, Vol. 35:3, pp. 295-319.
Hartley, R., Zisserman, A., 2000. Multiple View Geometry.
Cambridge Univ. Press, Cambridge, UK.
Another source of deviation from the homography flow is
distortion resulting from the camera (Michaelsen et al. 2004).
Particularly, cameras that scan the image using rotating mirrors
and only a few sensors exhibit strong non-projective distortions.
They violate the planarity assumption inherent in the pin-hole
camera model. More recent and future thermal cameras feature
focal plane array sensors and thus overcome the problem. The
example video of this contribution was obtained by such
modern device. The remaining non-projective lens distortions
are a minor problem. A linear radial symmetric model for this is
usually sufficient. In the estimation procedure outlined in Sect.
3 of this contribution such distortion estimation with one
parameter is included.
In Sect. 2.1 we excluded boundary locations from further
processing for this contribution. However, it is possible to use
straight lines instead of point locations for homography
estimation as well. Following this rationale the boundary
locations have to be connected and prolonged into sufficiently
long and straight line segments. This is going to be one of our
future research topics. Freeform boundaries can also be used in
this context following the approach of Akav et al. (2004). It is
obvious from Fig. 10 that this opens the way to utilize much
more of the information contained in such data.
Köthe U., 2003. Edge and Junction Detection with Improved
Structure Tensor. In: Michaelis B., Krell G. (eds.) Pattern
Recognition, DAGM 2003, LNCS 2781, Springer, Berlin, pp.
25-32.
Krüger, N., Felsberg, M., 2004, An explicit and compact coding
of geometric and structural image information applied to stereo
processing. Accepted for Pattern Recognition Letters.
Kraus, K., 1994, Photogrammetrie Band 1/2, Dümmler Verlag
Bonn, Germany.
Michaelsen, E., Kirchhof, M., Stilla, U., 2004. Sensor pose
inference from airborne videos by decomposing homography
estimates. In: Int. Archives of Photogrammetry, Remote Sensing
and Spatial Information Sciences, Vol. XXXV, Part B, XXth
ISPRS Congress, Commission 3, pp. 1-6.
Michaelsen, E., E., Kirchhof, M., Jaeger, K., Stilla, U., 2005.
classification of local structures in airborne thermal videos for
vehicle detection. In: Int. Archives of Photogrammetry, Remote
Sensing and Spatial Information Sciences, Vol. XXXVI, Part
8/W27, 3rd International Symposium Remote Sensing and Data
Fusion Over Urban Area URBAN2005
References
Akav, A., Zalmanson G.H., Doythser, Y., 2004. Linear Fature
Based Aerial Triangulation. In: Int. Archives of
Photogrammetry, Remote Sensing and Spatial Information
Sciences, Vol. XXXV, Part B, XXth ISPRS Congress,
Commission 3, pp. 7-12.
Dreschler, L., Nagel, H.-H., 1982. Volumetric model and
trajectory of a moving car derived from monocular TV frame
sequence of a street scene. CGIP, Vol. 20, pp. 199-228.
Ernst, I., Hetschler, M., Lehmann, S., Lippok, A., Ruhé, M.,
2005, Use Of GIS Methodology For Online Urban Traffic
Monitoring. In: Int. Archives of Photogrammetry, Remote
Sensing and Spatial Information Sciences, Vol. XXXVI, Part
8/W27, 3rd International Symposium Remote Sensing and Data
Fusion Over Urban Area URBAN2005
Faugeras, O., 1993. Three-Dimensional Computer Vision. MIT
Press, Cambridge, Mass.
Felsberg, M., Sommer, G., 2001 The monogenic signal. IEEE
Trans. On Signal Processing, Vol 49 (12), pp. 3136-3144.
68